What should a neural network diagram include?

At minimum, data flow direction, layer types, and the tensor shape at each stage. Skip connections and supervision points should be drawn explicitly. Training details such as optimizer and batch size belong in the text.

Should I draw individual neurons?

Only for teaching small examples. Research papers should use layer level diagrams, because neuron level drawings do not scale and obscure the architecture rather than explaining it.

How do I show tensor shapes clearly?

Annotate the output shape beside each stage in a consistent format such as 64×64×128. This lets readers verify dimensions without reading code and is the single most useful addition to an architecture figure.

What tool is best for neural network diagrams?

draw.io and Inkscape give full control for publication figures, Graphviz and D2 keep diagrams regeneratable, and PlotNeuralNet produces polished CNN visuals. Netron is best for verifying what your model actually contains.

How do I represent repeated blocks?

Draw the block once in detail, then show repetition with a multiplier such as ×6 or a stacked outline. Keep every instance visually identical so readers do not infer differences that do not exist.

Should architecture diagrams use colour?

Colour helps, but encode layer type by shape as well. Colour alone fails for colourblind readers and disappears in grayscale printing, so it should reinforce a distinction rather than carry it alone.

How to Draw Neural Network Diagrams

A good architecture diagram shows data flow, layer types, and tensor shapes.

Those three things are what a reader needs to reimplement your model, and they are exactly what most architecture scientific figures omit. A diagram of undifferentiated boxes joined by arrows looks like a neural network without communicating one.

A neural network architecture diagram with layer types, tensor shapes, and skip connections labeled (Figure generated with SciFig)

Decide the Level of Abstraction First

Three levels are common, and mixing them in one figure is the most frequent mistake.

Level	Shows	Use when
Neuron level	Individual units and weights	Teaching how a small network works
Layer level	Layers as blocks with shapes	Nearly all research papers
Block level	Repeated modules as single units	Deep architectures with repetition

For a paper, layer level is almost always right. Neuron-level diagrams do not scale past toy examples, and block level alone hides the detail a reader needs unless you expand the block separately.

See AI Scientific Figure Generation in Action

Watch how researchers create publication-ready scientific figures from text descriptions.

Explore the Tool

What to Label

Layer type. Conv, pooling, normalisation, attention, fully connected. Distinguish by shape or colour, and be consistent.
Tensor shape. Annotate the output shape of each stage. This is the single most useful addition, and it lets a reader verify dimensions without reading your code.
Key hyperparameters. Kernel size, stride, number of heads. Only where they are not obvious.
Skip and residual connections. Draw them explicitly. They are structural, not decorative.
Loss and supervision points. Where gradients enter, especially with auxiliary losses.

Leave out learning rate, optimizer, and batch size — those are training details and belong in the text or a table.

A comparison of an unlabeled box diagram and the same architecture annotated with layer types and tensor shapes (Figure generated with SciFig)

The same model drawn at neuron, layer, and block level to show which abstraction suits a paper (Figure generated with SciFig)

Conventions Readers Expect

Flow left to right, or top to bottom. Pick one and hold it across every figure in the paper.
Encode by both shape and colour. Colour alone fails for colourblind readers and in grayscale print.
Keep repeated blocks visually identical. If two blocks look different, readers assume they are different.
Show the input and output explicitly, with shapes. Many diagrams start at the first conv layer and leave the reader guessing what goes in.

Tools

Tool	Approach	Best for
draw.io / Inkscape	Manual	Full control, publication finishing
Graphviz / D2	Text-defined	Regeneratable, version-controlled
PlotNeuralNet	LaTeX	Polished 3D-style CNN figures
Netron	Automatic from model file	Verifying actual architecture

Netron deserves a special mention: it renders a diagram from an ONNX or similar model file, which means it shows what your model actually is rather than what you believe it to be. It is not a publication figure, but it is an excellent check before you draw one.

A workflow showing a model file inspected in Netron, then rebuilt as a clean publication figure (Figure generated with SciFig)

Tip

Draw the diagram from your model definition, not from memory or from an earlier draft of the paper. Architecture figures that disagree with the released code are a common and avoidable source of reproducibility complaints.

Create Scientific Figures Now

Describe your scientific figure in natural language — get publication-ready illustrations in minutes.

Try Free

The neural network diagram generator builds a labelled architecture diagram from a written description of your layers, and text to figure covers the other figures a methods section needs.

How to Draw Neural Network Diagrams

Decide the Level of Abstraction First

What to Label

Conventions Readers Expect

Tools

FAQ

Related scientific figure resources

Continue Reading

Block Diagram: How to Make One (2026)

How to Write a Conceptual Framework

Mapping Diagram: How to Make One (2026)

Publication-ready scientific figures, in minutes

Decide the Level of Abstraction First

What to Label

Conventions Readers Expect

Tools

Related Reading

FAQ

Related scientific figure resources

Continue Reading

Block Diagram: How to Make One (2026)

How to Write a Conceptual Framework

Mapping Diagram: How to Make One (2026)

Publication-ready scientific figures, in minutes